images-discussions

[dim cradle]

you're safe with fred durst around

[glossy scroll]

Lol

Fred's got my back

Now I think I'm gonna go make myself a pizza

...I could not live without pizzas...thank god for Italians.

[dim cradle]

aight, i'll make more scary folklore

[glossy scroll]

My fave.

[dim cradle]

The Wendigo - This creature is part of the folklore of the Algonquian people in North America. It's a malevolent spirit that possesses humans and turns them into cannibalistic monsters during harsh winters.

0Shot is going to town writing prompts for The Kuchisake-onna but I can already tell based on the horror folklore that moderation is gonna push back -- woah those markdown titles are huge lol

[shut niche]

These Paul Bunyans... 🤣

[dim cradle]

wow, paul bunyan reimagined

[shut niche]

Alright, time for...

[glossy scroll]

Wait!

[glossy scroll]

Paul Bunyan and Bison had a baby

[shut niche]

And that baby was born with HGH in it's veins instead of blood.

[glossy scroll]

Lol 💯

And they named him Bison Bunyan

"Lito BB" for short

[late blade]

Interesting made a prompt on a new chat DALL-E and suddenly it told me the prompt couldn't be done because I was referencing Final Fantasy games. I don't think a female version of Achilles in a Village full of Vampires where the Sky is filled with Shooting Stars and an UFO is roaming the sky is from any of those games. At least not to my knowledge.

[glossy scroll]

That's definitely Final Fantasy 5

[late blade]

😲

[glossy scroll]

And when I say definitely I may be exaggerating with nonsense

But who knows

[late blade]

I wonder what's happening

[glossy scroll]

They're probably working out the bugs with people trying to bypass ways to make ChatGPT produce things that it's not supposed to be producing.

[late blade]

there's something in the works for sure

[glossy scroll]

Could be that. You know they never tell us anything so anything is a guess.

We just have to make an educated guess.

I love guessing though...I once guessed this guy's entire life and I got it right.

[late blade]

That was just a test out of boredom. I should have said I wanted to go back to work in January and not February… 😡

[glossy scroll]

Yeah you should have

Topiadys

[late blade]

🧐

[glossy scroll]

Today I figured out something I've been researching for years

But now I'm confident it's the answer

But I can't tell you what it is

[late blade]

Oh no 😥

[glossy scroll]

Yeah it's a secret. It's in my profile though.

So you can guess maybe

[late blade]

hehe, no clue

[vapid elk]

I'll need to have a minotaur boss on the next dnd game jsut to make a cool token like that xD

[wispy storm]

I did a cat and mouse game with minotaur. I made labyrinth with dungeonai, created with dalle the minotor. Player have to roam around, hide in spot I made. It was on vttfoundry. Was fun

[vapid elk]

loved the idea

will use that xD

[wispy storm]

I should sell all my 5 years of riddle, puzzle, escape game, story, npc

[late blade]

Boo, I got timed out because the figure I wanted to depict and show in today's theme has a spanish name

Only used her name, so it got flagged as non-english

I just noticed I did a ton of images for today's theme and we got so much time for the day to end...

[wispy storm]

[late blade]

I feel threatened by an avocado....

If I say that to my colleagues, they will send me to the hospital...

[crystal bronze]

Hey guys is there any way i can use dall e to remix images?

for example if i give a photo of an object, can i ask it to generate a person holding it or something?

[final compass]

It can interpret what is approximately in the image and include a description of the object into the prompt.

[viral moth]

you can try with dalle-2 labs there is generate fill https://labs.openai.com/

the editor look something like that ig that waht you looking for

[earnest flame]

So dall-e 2 on labs.openai.com will generate variations, but it doesn't let you choose a prompt so you can't ask for it to change the image I'm some way.

[vagrant zodiac]

(and it uses dall-e 2, not 3)

[thick smelt]

i made the avocado version of the hotline bling meme

[hot rain]

Making the Blue Jellyfish Jelly Krabby Patty from Spongebob:

[hazy warren]

Is there a way to produce consistent images with DallE like Midjourney can do with some finagling? Would it be possible to automate this process without manually inputting prompts?

For instance, the same actors/characters for the purposes of storyboarding.

[late blade]

ok I'm drained

let's hope next theme is free cookies, muffins for everyone

Btw, I'm just uploading today's theme to this https://discord.com/channels/974519864045756446/1197549427624259624 if you guys have also something that won't make it to the theme, you are welcome to add it

Had so many pictures, impossible to post them all in the daily theme channel.

[green pebble]

Beeen working hard on tweaking up my bots to really push out the finest of quality

I think I am getting pretty close.

[median kraken]

👀 hello people, anybody has experience with the two different platforms for DallE use? ChatGPT and Copilot.

I am about to get a paid subscription, I'm trying to decide which one could be better for me.

[formal osprey]

Copilot is free in Bing afaik. ChatGPT offers more tools and is more competent overall.

[green pebble]

Yep there we go, pulling crazy details out of it now

[runic granite]

Attila

[median kraken]

thanks

[runic granite]

Copilot pro is sub based and has 100 boosts per day then slow gen for image gen 1 boost 4 images

anybody experience error with the image gen on GPT too?

[final jungle]

Hey folks, Have yall had much time to check out the gpt store for dalle? Had any luck with them? I feel like cultivating a prompt in gpt, then throwing it into dalle is still working better.

[ripe wraith]

I like this one so far https://chat.openai.com/g/g-B0bw75oD0-0shot

[late blade]

I got it on my TODO list

But Merlot has priority

[cloud dome]

Prompt?

[late blade]

Prompt: Can you make Merlot scared!

[cloud dome]

😦

[late blade]

lol

I'm telling the truth

[cloud dome]

So rude

[late blade]

lol

See, I don't lie, prompts are super easy

but Merlot GPT is not ready for testing yet

[runic granite]

Csodaszarvas

[late blade]

I had to look it up

[cloud dome]

What was the prompt for yours lol

[wispy storm]

I'm a simple man ;

Cute and angry menacing avocado, anime style

[cloud dome]

Wow. Did not expect that

[wispy storm]

[empty kelp]

A hyper realistic wide HD photo. On a dark night on a beach in Hawaii a fish stands on it’s tail, wielding two gleaming katanas with it’s fins. There are coconut trees. It is pouring rain and a powerful sea breeze blows everything

[wispy storm]

Finn diesel

[empty kelp]

it’s a sword fish

[wispy storm]

Finn diesel sword von fish

[cloud dome]

That's just weird

Literally doing the exact same prompt

[hoary garden]

GM everyone ! hope ur doing well

[late blade]

Good morning

[hoary garden]

where can i find logo makers ? or a prompt to use it creating a logo for my fb store page

[wispy storm]

Might be the CI ?

[cloud dome]

I just randomly decided to do a superhero cat and @late blade posted that

[wispy storm]

You can ask to dalle 3 , make a logo of x thing

[wispy storm]

Oh !

[hoary garden]

yea im doing it but like i suck with prompts lol

[dim cradle]

some impressive stuffs

[knotty wedge]

Is there a way to edit areas of a picture like in the previous editor? I want to change just certain areas

[plucky hare]

There's no inpainting feature in DALL·E 3 on ChatGPT, but you can inpaint using DALL·E 2 at labs.openai.com.

[late blade]

lol, it's so easy to make that

[dense mesa]

I'd library in this library

[late blade]

trying to find something interesting to do, been doing a challenge that each day a new workflow

[dense mesa]

make a new gpt by reverse engineering prompts

[late blade]

done that, next

o.o

all you need for my cat, my secrets are in there

[cloud dome]

[celest pebble]

can someone help me, I'm new at this and having problem. Let say you want to create an illustration flat style, where you have 5 judge in a panel behind a desk and they each hold a sign with different rating, how would you do that? What are the steps? I only get non sense stuff like all the judge have 1 star on their sign, or the stars are different colors, etc.

[wary stump]

Dont use the word rating. Just say they are holding up signs with single digit numbers

[dense mesa]

best you could do would maybe be describe each judge clearly on its own, but even then they'd probably blend together. don't think there's any kind of sure thing solution for your issue currently

[celest pebble]

anyone have tried copilot pro? I try the same prompt with copilot free and it seems to give me better result, just that I can't do 16:9 with copilot free

ha.... You've reached the current usage cap for GPT-4, please try again after 3:47 PM.

[dense mesa]

that message is really icing on the cake when you're frustrated, lol

[celest pebble]

yeah, its like it keep giving you things you don't want, and then, enough...

[dense mesa]

the main underlying issue is compelling gpt4 to pay attention to what you are instructing it to do

[late blade]

We need to enforce people saying hi....

I refuse to help people that don't say hi

[celest pebble]

Hi 🙂

[late blade]

Nice to meet you

[celest pebble]

sorry, was a bit frustrated with Dall-e

[late blade]

we all at some point

working with dall-e still takes good amount of work, so keep that in mind when you do a first prompt

[celest pebble]

what do you mean? I think I have something missing

I ask gpt4 turbo for help to ccreate the prompt

[late blade]

the prompt you did is short and leaves many things open

[celest pebble]

(but not the one in chatgpt, I don't want to waste one of the 40 request)

[glossy scroll]

I've updated my #1187233013956874260

[late blade]

try to leave as little interpretation room for dall-e, so be very thorough

for example the rating, you could change it to score, score from what range? 1 to 10, 1 to 5?

[celest pebble]

Create a 2D flat art illustration for a blog post header in a 16:9 aspect ratio. The image should show a line of stylized, abstract figures, each with a distinct, simple shape and a shade of blue. They are seated behind a sleek, modern table against a white background. In front of each figure is a card with a star rating, from one to five stars, in a variety of bright colors. This minimalist scene represents a panel of judges giving their ratings, emphasizing clarity and the theme of expert review.

That was one of the many prompt I tried

[late blade]

are there boards, whiteboards, lights?

is it wooden, is it plastic?

how are the judges aligned

minimalist illustration: what kind of line art, what color palette, are judges diverse or all are robots?

what is an expert review?

see, you can expand on your ideas and be more precise

hence I say, a good image takes work

[glossy scroll]

Thank you

Looks awesome

[late blade]

hehe, I knew you'd like it

[celest pebble]

I have tried judge, but it was doing like those judge at court

[late blade]

exactly, that's another thing, sports judge, beauty judge, muffins dressed as judges?

[celest pebble]

flat art illustration (its in there)

[glossy scroll]

Do it! 💯

[late blade]

yes, but flat illustration is anime? cartoon? ink? colored pencils? print?

try to leave as little chance to change stuff with your prompts

[celest pebble]

its a specific style: Flat art illustration, also known as flat design, is a style of graphic design that uses bold, bright colors and geometric shapes to create a visually appealing and clean look. It's characterized by its simplicity, minimalism, and lack of perspective.

[late blade]

yes, I understand what you are saying, but dall-e doesn't

[celest pebble]

that part is fine, it gets it

it's really the sign that I can't get right

[late blade]

ok, well, then you need to expand on that info that is missing

[celest pebble]

I even tried to specify how many stars and their alignment for each judge and it still showed 1 star or 3 star all misaligned

[glossy scroll]

Try using Image GPT Generator to make that specific style, in #1187233013956874260

[late blade]

that's another approach, Pythagoras made that really good guide

[glossy scroll]

Trying to make it easy for everyone to make images

[late blade]

just make sure you don't leave stuff to chance

[celest pebble]

how long are you prompt typically?

I'm afraid if its too long it will just summarize it before sending it to dalle

[late blade]

that was just for the art I wanted to use, but it was worth it

you can see https://discord.com/channels/974519864045756446/1197549427624259624 that the work I made were consistent with the images I posted there

[glossy scroll]

Usually the medium size prompt is best. As long as it details all the instructions first before making the image. That always helps Dall-E formulate a better image. Because it reviews its own description.

[celest pebble]

that's like 10 times the lenght of my prompts

[late blade]

hehe

[celest pebble]

I will go longer than and more details

[late blade]

you can see for example @open trench 's Scampers and how he consistently gets his project https://discord.com/channels/974519864045756446/1193219783877988423 forward

Don't be afraid to hit the image limit cap, if you see something you don't like, change it

dall-e won't go away

[celest pebble]

Im already at the usage limit

[late blade]

hehe, welcome to the club

[celest pebble]

Im thinking of subscribing to copilot pro

at least no limit

[late blade]

either way, with dall-e OAI or MS, just make sure you have a workflow

most people just do the lazy prompt and that's about it

[glossy scroll]

Yeah, lazy prompt no good

[celest pebble]

But so far I send sometime the same prompt to MS and sometime that one nail it, just that it's square picture so I can't use them

[late blade]

that's possible, sure

just be mindful, that it could change or you might need more specific stuff

[celest pebble]

I really didn't know that I prompt could be that long to be honest, I will start going more in details

[glossy scroll]

Mines only 6 inches long

[late blade]

you can also see #prompt-engineering if you want more information on how prompts work

[glossy scroll]

The prompt I mean.

I let Image GPT Generator do it's thing by giving it a small detailed instruction

I keep telling em Dys!

[late blade]

btw I didn't apply to the #spotlight post about mods and guide

[glossy scroll]

No you

[late blade]

one of my cat girls has 27 pages of prompts.....

I have to consolidate that somehow....

[glossy scroll]

Every bit helps

No, jk

It hallucinates after a while

People putting pages and pages and pages of prompts into a GPT only makes it worse

[late blade]

surprisingly with custom GPTs and knowledge you really got some cool stuff going

[glossy scroll]

Meee?

[celest pebble]

any GPT worth using that help?

[glossy scroll]

I keep telling you debian, use Image GPT Generator in #1187233013956874260

[late blade]

This one is a solid one https://chat.openai.com/g/g-B0bw75oD0-0shot

but it also depends on what you want to do

I love the Wolfram GPT for ADA, it's really good

[glossy scroll]

You can make your own custom image GPTs @late blade

Is that not what you want?

Is that not what everybody wants?

[late blade]

I am working on 2 custom ones, but haven't shared them yet

[glossy scroll]

Well if you do decide to share them, I look forward to them.

[late blade]

yeah, I still need to make sure I'm ok for beta

[glossy scroll]

Me 2

[celest pebble]

I'm at my usage limit, but I will try those GPTs later.

I hate the usage limit

I wish I could just run dall-e on my own computer

[late blade]

you gonna buy your own datacenter?

[celest pebble]

when did openai bought one?

how do you manage to put text on image?

[late blade]

Prompt: Let's do an image of Merlot with a speech bubble saying: "Hi, I'm Merlot!"

[celest pebble]

can you share the full prompt you used to create that? I'm curious to see how it looks like

[late blade]

that's the full prompt

Like this example: #images-discussions message

[empty kelp]

Geckos holding up number cards: 7, 5, 8, 9, 6. Please don’t change the prompt.

[late blade]

meeet @empty kelp he's also really good with prompting and coming with crazy ideas

[empty kelp]

DALL-E 3 isn’t able to connect sets of things, so you should only have one set

[celest pebble]

that's what I got... (with MS)

[vital gull]

usually the model doesn't have any problem depicting text if the text is short and using pretty common words

[celest pebble]

I tried generating a calendar page, it failed, they can get numbers in order

[late blade]

but yeah, what @vital gull says is 100% correct, short text is fine

[vital gull]

you may also get lucky and generate some phrases, but rare, and sentences, is still inconsistent

[empty kelp]

if you say “5 people holding 5 cards” it will draw 30 people. instead you need to list five cards, and let DALL-E attach a person to them

[vital gull]

it's a common issue the model doesn't "count" like we do

[celest pebble]

I'm gonina try that

[empty kelp]

it doesn’t do math, sequences, or order, but it can handle a short list of things

[late blade]

suprisingly with calculus it can do good explanations

but yeah, thea longer the text the harder it will get to make images like that

[celest pebble]

we ram out of time 😉

[late blade]

lol

I think next week I'll work more on my elven couple gpt

have to continue the story

[celest pebble]

goutot vartin

[modest sand]

what is the prompt?

[late blade]

let me look for it, one sec

[modest sand]

is there a prompt for web design prototypes?

[late blade]

it's possible

[late blade]

I didn't save the socks chat....

was looking for some success

[late blade]

if you really want to go into text tho, @grizzled loom is your person

[empty kelp]

ChatGPT 4 does say that DALL-E 3 can understand ordered sets and specific numbers of things — but so much random stuff is added to the prompts that only very simple things actually end up working

[plucky hare]

ChatGPT cannot always be taken as a reliable reporter of its own, or of DALL·E's, capabilities.

A bit more info about why this is a challenge for DALL·E can be found in its research paper:
https://cdn.openai.com/papers/dall-e-3.pdf

[vital gull]

yes, don't take its words as a must

[dim cradle]

haha love this one

anyone else get copilot pro? pretty cool the dalle image are now true hd. but i cant figure out how to make them square, is always widescreen

[shut niche]

Really? I haven't been following the news on Copilot. By true HD, what's the output resolution? Or do you mean they finally caught up, and are using 1792x1024?

[dim cradle]

yes i think. it jsut has the fidelity of gpt now.

they release a copilot pro option (20 usd) and it integrate into office and thing too

you get 100 boosted image a day

so far i am really loving it, feels nice and integrate into all my things.

[shut niche]

Sounds like things are about to get heated. MS office integrations with AI are probably going to get powerful IMO. 100 images isn't bad either. It's 200 with ChatGPT+, but people using MS Office are doing other things, so that seems like an appropriate #.

[dim cradle]

well and you always get 4 images from one token too. and even if you run out the 100, you still can gen images and with not much wait from what i have seen

but yes it seem a first big step into the ai being a real part of your computer

I have just start to use it after cancel my regret MJ 6.0, so i am sure there are lots of cool things to do with it i have not even scratch surface of

[empty kelp]

“a lineup of different colored cars in the order red, blue, green”

ChatGPT’s order examples work

[celest pebble]

So image gen is truely unlimited with copilot pro? They also mention that they assign more compute to each pictures to generate more details and better result, have you noticed any difference? I think I will just stop waisting my time with this and order copilot

[dim cradle]

I jsut start using it today so I am no expert yet. I am pleased though is all i can see. images look great but of course its dalle.

[plucky hare]

You might try regenerating a few times (if you feel like experimenting!), or trying a higher # of objects -- I wasn't replying to say that DALLE can never do this type of thing ever, but just that it's inconsistent, limited, and not wholly reliable! But that's no judgment from me, it's still an amazingly capable model.

[dim cradle]

one nice small bonus for image gen on copilot pro is you can make copyright stuffs since they are more relax about that kind of thing that gpt i guess

[empty kelp]

a lineup of different colored geckos in the order red, blue, green, yellow, purple. the geckos hold a different numbered card over their head in the order 1, 2, 3, 4, 5. please don’t modify the prompt

[dim cradle]

we have to retake the photo, yellow gecko has one eye closed.

or maybe he is just winking haha

[empty kelp]

it got the colors here. i wonder if there is a way to describe this so the colors and numbers are in order

[celest pebble]

@la3 can you try that prompt on copilot pro? Geckos holding up number cards: 7, 5, 8, 9, 6. Please don’t change the prompt.

[dim cradle]

sure i dont think it will be much different since it is still dalle3

[plucky hare]

I'm sure with enough attempts it would eventually get both right!

[empty kelp]

you always need to tell it to not modify the prompt. it will still scramble the prompt a little, but if you don’t say that it totally scrambles it

[dim cradle]

you dont like Geckos?

[empty kelp]

you need to say the gecko is standing on it’s hind legs for it to hold the card. i forgot to put that in

[celest pebble]

I don't think a company have a sexual orientation

[verbal sequoia]

  Attention  Off-topic chats have been cleared by a moderator.

[glossy scroll]

Boom

[celest pebble]

lets get back on business and talk about dalle 😛

You've reached the current usage cap for GPT-4, please try again after 6:35 PM again...

[glossy scroll]

Awwww

The timing is terrible

It happens when you least expect it

[shut niche]

I don't know why... but I feel like you should make a gecko xylophone now, ha ha. Something about those colors. 🙂

[sullen fox]

squidman

[glossy scroll]

I like gecko's

[dim cradle]

i like turtles

[dim cradle]

the api has a quality setting that accepts standard or hd. bing has always been using standard.

i don't have any rhythm. i can barely spell it.

[earnest flame]

I'm going to save this snippet so I can bring it up every 50 times a month when someone wonders why an image ask is challenging.

[dim cradle]

The OpenAI Cookbook has a figure showing the difference between standard and HD, even though it's the same resolution

[shut niche]

When he said 'true HD' I was just looking for some extra resolution information, in case Copilot had added something different than 1792x1024. HD tends to mean different things in different applications, lol. Companies often "inflate" the term HD IMO, lol.

[dim cradle]

Ah yeah, it hasn't gone 4K yet!

[glossy scroll]

I like squid's too

[sullen fox]

nice

[dim cradle]

imagine when dall-e goes 4K though, wow.

[shut niche]

I'm curious if OAI will bump Dall-E to 4K, while releasing a video model that's similar to the current Dall-E3 (or what resolution that it'll be if and when they do). Too bad we're always "speculating." I'd love some little hints as to what's coming. Wink wink OAI 😜

[shut niche]

I really like that one, that's a really cool concept.

[dim cradle]

nice. i was going to say maybe say in prompt how many cards or geckos there are to start

maybe they will release 4k with some video when gemini ultra is release.

because i think ultra is rumor to have some things like that. they will need to compete

[glossy scroll]

Gnomes get around

[empty kelp]

the key is to not say how many geckos or cards there are. if you say that it draws 20 of each. it doesn’t grok the numbers

well, i did in that one… but i think it works better if you leave out the number

[dim cradle]

hmm i was going to say you just prove otherwise but ok haha

[dim cradle]

i still don't see how they can reasonably release video animation to the masses -- that's 30-60 frames or more per second, and they can barely handle the demand for a single still frame.

true. and most video seem far away anyway. everything i see is 2 seconds of jittery or scanning across nothing impressive really yet though i am sure dalle will of course be able to top this

[celest pebble]

We get 40 prompt every 3 hours, at 60fps, you could generate 1.3 seconds of video every 3 hours, and if you don’t like the output good luck

[dim cradle]

in a few years maybe

it will be impressive. we need more power i guess haha

[empty kelp]

There are 4 cards (white backgrounds and black text) with numbers in the exact order 1, 2, 3, 4. There are 4 geckos with colors in the exact order red, green, blue, yellow. Each gecko stands on its hind legs and holds a card over its head with its front hands: red has 1, green has 2, blue has 3, yellow has 4. Please don’t modify the prompt.

the colors are red, gren, blue, and blow

[glossy scroll]

4 blow?

Oh man

[dim cradle]

ya something to look forward to 🙂 I have every confidence in OpenAI's ability to circumvent the technical challenges, they have a rather proven record already. Altman says human-tier AI is coming this year, so I wouldn't bet against them by any means.

[unreal hill]

I'm a novice to generating images with dall-e. Where are the best resources for learning how to write good prompts?

[quartz bronze]

Getting a white screen when i log in, any solutions?

[glossy scroll]

#1187233013956874260

[unreal hill]

thanks!

[celest pebble]

You've hit your daily maximum number of images. To ensure the best experience for everyone, we have rate limits in place. Please wait for the next day before generating more images. Your daily maximum will reset in 15 hours and 44 minutes. 😭

[quartz bronze]

I haven't log-in in like a month.

[glossy scroll]

My pleasure

[celest pebble]

Dall-e hates me

[dim cradle]

maybe try another support channel in this discord, this is for discussing dall-e

[quartz bronze]

this is about dall-e

[dim cradle]

yeah but that doesn't include login issues

[glossy scroll]

That's more for like #chatgpt-discussions

Or #gpt-models

[sullen fox]

i don't think they're having issues, i think they're just saying they haven't logged in

[dim cradle]

you can ask question in here too, this room has some of the most dedicate dalle3 users always sharing tips and thing

[quartz bronze]

its not a login issue, whenever i try to generate an image the screen goes blank.

[glossy scroll]

Maybe you need to ask Dall-E more politely

[dim cradle]

browser issue maybe 🤷

[glossy scroll]

Probably yeah

Cache issues

[quartz bronze]

i thought so, i cleared cookies, allowed java-script, tried in firefox, chrome, and on my phone. same thing keeps happening.

[glossy scroll]

have you tried logging out and in?

[quartz bronze]

yep

[dim cradle]

using gpt4 or labs?

[quartz bronze]

labs

[glossy scroll]

Yeah, those are questions we don't know. Try reporting it #1070006915414900886 and ask in #community-help .

[unreal hill]

What is the differencee between asking dall-e directly for an image and using the Image GPT Generator I found that you shared @glossy scroll #1187233013956874260 message

[glossy scroll]

So, the Image GPT Generator will just help you make a specialized Image GPT for any specific style you want, or any universal style as well. That's if you don't want to go into the nitty gritty of it all and just want to make an Image GPT.

The difference for asking Dall-E directly - means that if you want images to respond more to your specific requests, you make sure that you put things like Put these instructions in front of every image prompt you write, before you write anything else: then open up a quotation marks and say whatever you want and close it in quotation marks. That will basically embed your instructions within the output prompt when Dall-E produces images. Making sure that Dall-E reads specifically from the output prompt and not just let ChatGPT transform your words into a more refined language before pasting it into the output prompt.

The output prompt is the circle with the exclamation mark in it. When you click that, it opens up Dall-E's output prompt for the image.

That's the place Dall-E reads everything to create the image.

[unreal hill]

Thanks for the info. Still trying to wrap my head around this. I used your GPT generator and changed my custom instructions to fit the response that was given after I described my picture. What about the section in instructions that asks How would you like ChatGPT to respond? What typically goes there?

[dim cradle]

today's daily posts are all so good i don't know where to begin--so many ideas

[glossy scroll]

Noo, that's the custom instructions place. Go to 'Create GPT' then go to the 'Configuration' menu, then put the info in the 'Instructions' section.

The Custom Instructions place is specifically to guide how you want ChatGPT to respond with All GPT responses.

Think of it like tailoring your ChatGPT to give outputs based on what you want for everything, not just images.

The 'Create GPT' is an entirely other thing.

You need to be a GPT Plus Member

Btw

To have access to all that.

Except for Custom Instructions

[unreal hill]

Yeah I was using it for tailoring responses for code refactoring. I have the $20 a month plan. I've never used the create gpt before.

[glossy scroll]

Creating GPT's is where the Real Fun begins!

[unreal hill]

I can only imagine. Maybe I can use it in my website? I was just trying to get some good images for my landing page but now Im curious.

[dim cradle]

@shut niche your gpt is so good for reimagining variations.

[glossy scroll]

Yeah, my guide is meant for creating your own custom image gpt's, and playing around with the instructions section of custom gpt's.

It's farrrrr more powerful than just using regular stuff. In my opinion.

@unreal hill If you want to find different ways to play around with prompts and make Custom Instructions, just to get your feet wet, you should check out #1019652163640762428

They got some pretty good stuff in there.

[dim cradle]

yup, there's off-the-shelf and then there's the expertise that can truly level up your outputs

[glossy scroll]

yeah, I've learned so much from prompt labs that have helped me with my custom gpt's.

[unreal hill]

So I can make a gpt specifically for the code that I'm using for my website. And I can also make one specifically for creating the images for it?

[glossy scroll]

Yep

[unreal hill]

yeah I need to explore this more.. thanks for the info once again!

[glossy scroll]

My pleasure. Any time

[glossy scroll]

[dim cradle]

Custom GPTs are so powerful. This democratization of AI enables so much potential to help others and help yourself.

[glossy scroll]

Indeed they are

[vital gull]

how are you guys!!

[dim cradle]

Just enjoying life as much as possible with ai and art.

[glossy scroll]

Good. Just trying to make more GPTs! I'm addicted.
How about you?

[vital gull]

i'm exploring perspectives these days and how well the model understands, and where it struggles 🫡

[glossy scroll]

Ah, a trial and error way of life.

Admirable.

I can't wait for GPT5

They need to make my GPT making experience easier

[dim cradle]

gpt5 and dalle4 🙏

[dim cradle]

there have been suggestions there won't be a gpt5, rather a next-generation human-tier AI with a new name altogether.

[glossy scroll]

That would be nice.

I hope that's true.

[shut niche]

Thanks Shon! The positive feedback means a lot, especially since I've been working so hard on it since GPTs were first announced. I get really pumped seeing how everyone is using it.
I also hope it's helpful for 'noobies' that come in here, that don't care about learning prompt engineering, and just want great images from Dall-E.

[glossy scroll]

Lol

Ok

[shut niche]

I have a trick that, that sometimes works for unusual perspectives.
If you get it to mimic action camera extreme sports, you can evoke cool camera angles that are, otherwise, hard to get. There's obviously a million approaches, but that's a fun one to get lost. 😀

[cyan loom]

Why is /daily not active right now

[vital gull]

I use photographers approach, and architecture perspectives, they work pretty fine

[shut niche]

I just suggested it because Dall-E is so good at blending concepts, sometimes it does cool 'unexpected' things when blending action shots with stuff like architecture. 100%, a photographer's approach is great too.

[vital gull]

but the thing is usually regular users, who aren't very familiar with 'technical terms' they try to explain with their own words, os i am trying to learn how well model understands perspectives that are put in simple words.

[shut niche]

Gotcha

[vital gull]

e.g. how someone who aren't familiar with e.g. bird eye view, would explain it with their own words, and if model understands it

[dim cradle]

The same Rhythm of the Night concept -- requested a closeup of a dancer with a shallow depth of field.

[shut niche]

@vital gull That's something I'm ALWAYS considering with that GPT I've been working on. For instance, Shon suggested shallow depth of field yesterday. I could add that in, but then I need to explain to ChatGPT when and when not to use it.
But, if it just 'understood', then the user experience would be better, for the inexperienced who just want great images. They want to say, "Make me a cool image of X." Then don't think to add in all of the key concepts that everyone here would.
So I agree with what you're doing 100%. I've spent 'too much' time trying to figure out what Dall-E responds to (or doesn't). LOL

[cyan loom]

please help

[shut niche]

Maybe because I'm not posting a million images? LOLz (I'm allowed to poke fun at myself.)
I've been busy running gens so I can demo this GPT's abilities, and haven't done anything for the Daily today.
Feel free to check it out. 🙂
https://chat.openai.com/g/g-B0bw75oD0-0shot

[shut niche]

What do you need help with?

[dim cradle]

it's temporarily out of order

[shut niche]

OHH, ha ha

Did not know.

[cyan loom]

The /daily function

Damn , do we know when it will be back on? I want credits to try the dalle thing lol

[vital gull]

what do you mean doesn't work

[cyan loom]

I tried to do /daily (email) and it did not give me my 4 credits

[dim cradle]

we got a report some weeks back that it was broken so to speak, that's all i know

[vital gull]

that's great!! these models give you so much freedom to explore and iterate on new stuff, quite different what we used to with D2, so since that time my goal has always been "using simple minimum prompts" to test how the model would behave. i can imagine how frustrating it is when you have a vision in your mind and you don't have enough vocabulary or mediums to describe it, so im trying to explore alternatives, associations and etc

[empty kelp]

this is what they’re talking about

[vital gull]

yes I know, i just wanted to clarify what didn't work for them exactly

[shut niche]

Couldn't agree more. I've spent considerable time experimenting with CoT and ToT type mechanisms, so ChatGPT could natively just 'understand' a user's request better, because it IS undoubtedly frustrating when you lack the vocabulary, or expertise, to describe your idea. Just like you mentioning the endless possibilities in photography you might apply.
So I think the user's experience is 'generally' better if some of that is just 'understood' by GPT4 or Dall-E, and done for them.
The double edged sword is... if too much is automated, then it would all become bland or predictable, with the same techniques always being assumed and applied to 'said' styles.
But, GPT4's imaginary '3D worldview' and understanding of these art concepts is amazing when it's triggered to use them correctly, that's for certain.

That's always on my mind, not "adding too much" that I would remove peoples creativity, but adding enough that it 'boosts' their images. It's a challenging balance.

[glossy scroll]

That's why you create a GPT for that.

To help with the vocabulary.

I run three GPT's when creating one GPT to help me make my instructions.

All because I'm copying and pasting ideas here and there.

[vital gull]

got your point, but a regular user, unless they use someone elses' gpt, i don't think they spend too much time on working on their own, they just want something fast and quick and the model to just get what they mean, even if they lack the ability to properly describe their vision

[shut niche]

I both agree and disagree. I think in the short term, lots of specialized GPTs are needed. In the long term, more 'all inclusive' GPTs will be the way.
Think of all of the apps on your cellphone, and how annoying it is to have a different app for everything. I don't think a GPT for every niche' is the right path.

[empty kelp]

these GPT models have pretty spectacular reasoning skills already. when the processing time & cost is increased you can tell it semi-coherent things that barely make sense and it still figures out what you’re trying to say

[shut niche]

I agree with you across the board. I just meant, I've been working to do that w/ this GPT. We're 100% in agreement.

[vital gull]

and that's great! if there are users who can simplify the game for other users and they are willing to do so, why not, amazing!

[shut niche]

Bahahaha I ❤️ this. This just popped out of a test I was just running. Hysterical!

[glossy scroll]

I think it's a matter of learning to some aspect as well. Sure, you can try one thing and be happy with that. But if you want to make something more to your fitting, you have to make Custom GPTs. I've made three GPT makers for making Custom GPT's, to make it more accessible in understanding how to create Custom GPT's. But I guess it's a matter of preference, and how much you're willing to invest in learning these things.

#1172301858300235846

#1180661996811714580

#1197683008031952986

[vital gull]

I don't want anyone to misunderstand me, but the best method I've found for learning how to describe things accurately is to imagine that you're explaining it to someone who can't see. For example, in this case, it can be interpreted as describing the art style, illustration, or painting for a blind or visually impaired individual. The approach is very consistent and persistent in focusing on details, which is crucial for the average user who can only explain what they see without using technical terms or referring to specific mediums, artists, or associations.

[sullen fox]

“movie poster for Ice Age 6, but hyper realistic”

[glossy scroll]

Well then you would have to just use other people's GPT or wait for OpenAI to give you a more fluid All-In-One app. But even blind people have to adapt, and learn to read using the senses that they have. So blind people aren't helpless. Everyone has to learn at some point or another.

[vital gull]

what i meant is, uploading a picture and asking it to describe it for someone who can't see, then change the subjects in the prompt and recreate it, you don't have to use other people's GPT for that. I agree on the last part 🫡

[glossy scroll]

What do you mean by that?

Uploading a picture and asking it to describe it to someone. It doesn't do that already? I thought it did.

Me confused.

[vital gull]

uhm, e.g. I really like this illustration but i am not sure how to describe it to the model so it will give me a similar kind of illustration but with my own subject, i would ask it to describe it for someone who can't see, to not violate the copyright, and then ask to give me the verbal prompt, i would just edit the subject afterwards

[glossy scroll]

[dense mesa]

you can ask it to describe images in all sorts of crazy, novel ways

I've had it describe images in a way that it would convey to a blind person what it was looking at

[glossy scroll]

Why does the blind person need an image? Why don't you just have the GPT explain a scene instead of an image?

I am so confused, lol

[dense mesa]

I normally ask it to describe the image in a way that someone that never saw it could see it in their mind using the description

[empty kelp]

i felt like i was talking to a blind person when i tried to describe the geckos with colors and numbers to DALL-E

[dense mesa]

the blind description is cool for weird visuals, but doesn't exactly work out as far as looking similar

[glossy scroll]

Yeah

It's not the same as having the gpt interpret how to create the art vs describing it

It won't be as detailed or accurate

[green pebble]

Anyone have thoughts how I can make this use the whole 16:9 canvas?

[glossy scroll]

And that's because the GPT needs details in how to form the art, not how to visually explain it. Such as it's different for drawing a table than looking at a table.

[dense mesa]

try things like "expansive background"

[green pebble]

Ok I'll give it a shot

[dense mesa]

although might be a bit tricky with isometric type images

[glossy scroll]

Yeah, you might need to just crop it.

[green pebble]

Nope it worked very well thank you

[empty kelp]

you need to either specify a background, or describe the light as reaching a wider area

[glossy scroll]

Oh ok, that's cool!

[green pebble]

Yea the expansive background term worked super well thank you

[empty kelp]

you could also say, “the diorama isn’t in a box”

[vital gull]

landscape perspective also works

[green pebble]

I will try these

[empty kelp]

it had landscape perspective, but dioramas are typically in a box

[green pebble]

Like I like this viewpoint

But it just stops at the edges and I want it to just keep going if that makes sense

[dense mesa]

could try things like "medium view" maybe? normally applies to images of people, but might work with what you're doing

[vital gull]

you mean isometric perspective but in continuum?

[green pebble]

Yes

[empty kelp]

you could say it’s isometric and leave out the diorama part. or say diorama that’s not in a box

[vital gull]

don't mention diorama

[green pebble]

Hard to achieve currently, but I'm going to try the tips that have been shared and see potential results

[empty kelp]

there are dioramas not in boxes, but DALL-E leans toward boxes

[dense mesa]

don't think there's a single magic ticket for what you want. but if you incorporate enough concepts and hit the right seed or gen_ids it should work out

[green pebble]

With water it does it near perfectly

[dense mesa]

ask it how to do what you want it to do. doesn't always work, but sometimes it'll give you the answer

[green pebble]

Yea I tried having another GPT describe how it made this image lol

[empty kelp]

the ocean will blend into it if you don’t say it’s in the background. instead you can say, “the image has an ocean”

[green pebble]

Got it closer

[formal osprey]

That works relatively well. I literally used that sentence in the instructions for a GPT: please describe the provided image as if you were conveying its essence to someone who is blind.

In the end, it was not producing the most artistic results, it was summarizing the important elements but ignoring the details. Here's the final prompt for reference. But it's a complex task. Very tricky.

please provide an exceptionally detailed description of the image using precise technical artistic terminology and references to rendering techniques or shaders when applicable. Focus on elucidating the most salient elements, including but not limited to precise colors, materials, lighting, shading, textures, geometric shapes, and spatial relationships within the image. Employ your superhuman-level analysis to offer a comprehensive, deterministic understanding of the image's content and artistic nuances, ensuring that every aspect is meticulously conveyed. Optimize and minify the output as a 400 tokens DALL·E prompt encapsulated in a codeblock. No output is produced outside of the codeblock.```

(note, this is experimental but maybe it can inspire some people)

[shut niche]

The reason that Dall-E even works is because GPT3 has a similar imaginary worldview, intermixed with a more factual listing of objects from the ML synthetic data descriptions used to train it. It's 360 worldview allows it to understand complex concepts in a nuanced way, as well as the interactions between those concepts.
The ML synthetic data, on the other hand, ties it's knowledge in a less abstract way, to a more physical presence or naming of objects.
So I have to agree with Neighbor, explaining it like you would to someone that's blind is calling on it's worldview of concepts, or "worldly understand." If you're drawing a car, it comes with a steering wheel, and the model doesn't need to list, "Draw a car with a steering wheel."
Where a more factual representation helps, is when trying to mimic the language that was used to label the images in the synthetic data. Liked adding Impasto to oil, where you don't explain what Impasto is (which you'd need to do for a blind person).
E.g, the answer is both approaches are important.

[vital gull]

you can ask simply, visually impared persona, focus on important key elements, and it will understand what you mean

don't forget each data is stored with labels on it, back in d2 days, we were generating traditional art illustrations, in the same way it is labeled in museums, e.g. subject, artist, time, era, and movement. never missed

[green pebble]

Apparently pixel art just has the limitation because of all the training data as realism landscape bot I made can do the isometric perspective just fine

[shut niche]

Exactly. I try to call on both, concept, and cataloged name. I just wish we had the ML data so I could data-mine it for those specific naming details. It would be extremely handy.

[formal osprey]

My english is not the best, when I test stuff I usually use quite silly sounding language. But usually I ask chatGPT 3.5 to propose some variations and rephrasing, to test if it works better

I'm amazed that it understands what I mean at all often

[vital gull]

read both papers for 2 and d3 maybe it will help you understand how generally labels are put, but again sources are important, without sources you wouldn't guess how it is labeled, but again it's just my guess

[shut niche]

I read them. I mean, I wish I had ALL the data, lolz

So I could identify my own trends visually, by looking through all of the descriptions used.

[shut niche]

YES

[shut niche]

Yesterday I made a list of "power words," but for strong language commands.

[empty kelp]

it has a concept of views, viewpoints, and backgrounds. like here the image is split into two views, each view has a wave and vortex in the background, but the character is just in the image. that’s how it is with the diorama — you can put something into the diorama, the background of the diorama, the image, or the background of the image

[vital gull]

cool!! thesaurus is the way!

[vapid elk]

draw a wide image of a sign language chart 😆

[formal osprey]

I wish I could vary my fingers count on the fly sometime, to use sign language

[unreal hill]

that looks awesome.

[shut niche]

Ouch burn. Lol
I've been experimenting with which words carry more weight, so a thesaurus only gives you words to try. That's not = words that work.

[vital gull]

synonyms 👀

[shut niche]

Basically it's 100 and 1 ways to say, "do it or die!" To ensure commands are consistently respected. The fact that each response starts with randomness means it often answers the same questions differently. So in my experience, certain strong vocabulary helps to ensure commands are adhered to more universally. Upping the # of correct responses.

[vital gull]

ooh, i thought you were referring to something else..

[formal osprey]

I used to experiment a lot with prompts that encourage chatGPT to "try again after a mistake" (more prompt engineered than this 😅)

[vital gull]

rather than that, copy the prompt (of that image) and iterate on it

you will have more consistent output

unless you're trying to get different variants heh

[shut niche]

Trouble is, it makes its decisions early on, and new information or contradictory information aren't weighted as highly. It creates a domino affect downstream, by starting on "the wrong foot."

[vapid elk]

the Dalle custom GPT by OAI can make up to 5 images in one message, it is good for iterating, you can tell it to make 5 slight prompt variations

it stil lgenerates them one by one

[formal osprey]

if there wasn't a messages limit per 3h, I would still do that 😅

[green pebble]

Thanks, if you search up "Wright's Pixel of Wright's Sprite, or Wright's Landscape, or Wright's Statues" I built customized bots for art generation to go with specific tasks and generations

They have provided imense quality in various styles

For instance this one is just my traditional pixel bot generatior and it can handle some serious stylization

[shut niche]

You should link them. Sounds cool! 😎

[formal osprey]

I experimented with "role playing" prompts such as "the clumsy student making mistake and the patient teacher explaining the solution". but I admit this is more useful for code interpreter. less for art. I use a different approach for art prompts

[green pebble]

https://chat.openai.com/g/g-tXg5ekQNa-wright-s-pixel-painter-pro

There is the pixel art one, I'll get the other links and put them in my discord profile

[shut niche]

Yeah, the "act like an expert" works "more consistently" after the fact, like say, when it's writing an oil painting prompt like an oil painter would.
But it's less effective for solving puzzles.

[shut niche]

I'll check it out!

[green pebble]

Yea thats really fricking cool!

[shut niche]

Nice work man. 🙂 I hope it does well in the store!

[empty kelp]

GPT 4 creates an English prompt, and then DALL-E 3’s language transformer model converts it into a hierarchy of vector sets the diffusion model can draw. So you can group everything into a diorama or view, or put the diorama into a crystal ball, or a vortex into the clothes; etc.

[green pebble]

Thanks making a post in dalle 3 gallery to officially promote the bots

[shut niche]

@green pebble Not sure if you saw, I finally published the one I've been working on, since before we were all making "GPT dragons" in here, haha
You'll have to check it out.
https://chat.openai.com/g/g-B0bw75oD0-0shot

[green pebble]

I'll go look now

[shut niche]

Yup, concepts, objects, and relationships.

[green pebble]

For a super vague prompt, it handled creativity well

[shut niche]

Thanks! You picked up on that right away. I'm pushing it hard in that way, hyper-embellishing. There's a lot more, but that's a major facet.

[green pebble]

I'm going to try a specific prompt this time

[shut niche]

I was gunning to make something that would be great for everyone. If you're advanced, you'll still get what you want, just more of it. But if you're new to Dall-E, you don't have to spend hours learning how to manipulate ChatGPT or craft prompts, it "just does it."

[shut niche]

FYI, it'll give you 4 images per request if you ask it to.

[green pebble]

Interesting

[shut niche]

You can ask for 4 variations, and you'll get all sorts of amazing results. Or you can say I want 2 like this, and I want you to choose the other 2 for me.

[glossy scroll]

In all honesty, I still don't understand what Neighbor was referring to when explaining things to the blind and what you're saying corresponding to that. But I'm gonna give up on trying to figure it out, because I don't want to comment on something that I don't understand.

[shut niche]

Initially, it was a reference to the vision model describing all of the necessary details of an image "as a blind person would need them (holistically)", which captures the details in a useful way for it to be turned into an image prompt, negating the need for specific GPTs, when you can simply impersonate the style using an example image.
VS, someone trying to "explain" a style themselves, to ChatGPT.

It's the same as how the ML synthetic descriptions were made. It's just a vision bot reading and describing images in hyper detail "as if describing them to a blind person."

[glossy scroll]

But doesn't the default GPT explain the images already?

How is a blind person not getting info compared to a not blind person when they can hear the text?

[shut niche]

It boosts its description in a particular way. It's also why @dim cradle was having success using a different highly descriptive GPT I made to reverse engineer images for vision to dall-e generations. You can boost that ability, it's not always the same output from ChatGPT.

[dim cradle]

I wonder if there's an instruction for the same output that doesn't reference people with disabilities. For the equivalent, GPT-4 says, "you could ask for the image to be described in a highly detailed and vivid manner, emphasizing sensory details and spatial relationships to create a clear and comprehensive mental picture" which should have the same effect? in theory, havne't tested it.

[hearty ether]

Oh got any examples?

[glossy scroll]

Oh ok. Now that makes sense. Just a way to enhance an explanation for images. That blind thing threw me way off, lol

[shut niche]

Not offhand, but you can make one quickly. Just use Dall-E GPT or ChatGPT, and give it an image, and ask it to describe it.
Then, in a different session, give it the same image, and ask it to describe it like you would if describing it to a blind person.

Or, likewise, describe it meticulously, in hyper detail, making sure to include these attributes (placing whatever you want it to consider here).

[glossy scroll]

Lol. It would produce exactly that, visual representations of other senses, which would be blurs and lines and symbols.

It's quite literal.

There's no transcendence of AI.

It wouldn't be something miraculous.

[empty kelp]

one big difference between describing a scene to a person and DALL-E is that DALL-E has a set amount of processing time for your image. And things are actually grouped into sets. Things you tell it to focus on get more processing and interaction, and things you tell it to put in the background get less. And the more stuff that’s in the scene the more spread out your processing gets.

Like doing a handstands or balancing things takes serious processing, and if you don’t make it part of the focus it’s just not going to happen.

And if you want the ocean waves to affect a diorama it needs to be similarly grouped

[shut niche]

I would refer to what you're saying as "weights". As you add weight to any one concept, you remove weight from another, because the total % must total 100, so to speak.

[empty kelp]

if you try to put anything else into your scene when it’s trying to do something with complex support and balance (like this), all the characters will collapse on the ground

[shut niche]

I would hate to break their high degree of mindfulness and intense focus.

[dim cradle]

I ran some tests. Compared to the default vision description, requesting a description for a person who is blind did yield better results as far as reproduction of the same image; however, another test revealed that so did simply asking for a prompt that would reproduce the image.

[shut niche]

Ask it for an extremely verbose description of the image.

[empty kelp]

yes, you don’t want to do that. they’re all insane

[glossy scroll]

It's the same as having a GPT detail a description before creating the image.

Explaining an image and having to describe its detail before making it fall under the same functionality.

You'd have better luck trying to make GPT's that better explain an image before it processes it's art.

[shut niche]

You're right. Any extra emphasis on a hyper description should be about the same thing.

[glossy scroll]

No, what I'm saying is that if you made a GPT that made detailed images vs making a GPT that explains an image based on its details, they both fall under the same category. Explaining an image before producing it is the same in all regards. It doesn't matter if it's for an image describer or a custom image GPT.

[dim cradle]

Just curious, if you used separation of concerns pattern to group a more complex background, that foreground would still collapse quite literally?

I think there’s something to your observation that there is limited processing time allocated to an image generator so what you’re saying makes sense. I just like to know the technical details so I can be sure I’m relaying accurate info.

If we can find a workaround to that current limitation, that’s even better.

[empty kelp]

if you add a little more there would be lots of anomalies with three legs, heads on backwards… and then they can’t stack. it has to figure out precise limb, hand, and foot positions to balance everything — and it burns through all of the processing

[dim cradle]

yeah, if I'm trying to choreograph a dance scene, there is greater chance for anomalies given the more complex depictions of various subjects -- like i don't know if that dude on the right has 3 legs or if somebody is passed out on top of him lol

[empty kelp]

That’s why they have ChatGPT 4 as a middle man for DALL-E 3. it’s trained to focus on important detail, and reword it to lower the processing for unimportant things.

[dense mesa]

Gpt4 also says stuff like, "and now for our second image..." in prompts. It's a bit iffy as far as it's prompting skills at times

[dim cradle]

I'm sharing 4 photos for this vision model test. 1st is the original. 2nd is a reproduction using the default response. As you can see, it's not very good at recreating it.

[dense mesa]

That's because it can only convey so much in a prompt

I think it did a decent job actually

[dim cradle]

3rd is a request for a description for a blind person. 4th is a request for a prompt to duplicate the image. Both yield better results, pretty comparable to me, I don't see where one is better than the other -- and it's still missing some obvious details, namely that one tower stands above all the rest.

[empty kelp]

it scrambles it to make it safe and not be an intellectual property issue

[dim cradle]

this is a follow-up on an earlier discussion about the vision model and how to get the sensory and spatial details we want.

[dense mesa]

Gotcha. I have a got I was working in that was an attempt to really get image layouts down

[dim cradle]

the vision model is amazing

[empty kelp]

i’ll show you the main thing that ChatGPT 4 does for DALL-E 3

[dense mesa]

Explicitly describing things like orientation and placement. Not "behind the first man is a second man"

[empty kelp]

“””The front view shows the elf’s detailed facial features and the full front design of their ornate clothing. The back view focuses on the hairstyle from behind and the intricate details of the clothing’s back. The left and right side views display the profile of the elf, capturing the symmetry and differences in the hairstyle and attire from each side. Each frame is set against a simple, unobtrusive background to emphasize the character.”””

ChatGPT wrote this. You can see that it puts “focus” on things

And it tells it to not put detail into the background so more processing can go into drawing the character.

That’s why you have to tell it what to focus on — And otherwise it decides for you

It mostly just sorts your prompt into important and not important, and makes it so DALL-E can draw it with the processing time that you’re allocated (and it scrambles things a little to make it safe)

[dim cradle]

well, they're rapidly breaking through constraints so i'm sure we'll get there

[glossy scroll]

It's using its latest recompopulated scrambler

Cutting Edge AI technology

[empty kelp]

it’s all pretty amazing. you just have to put “Please don’t modify the prompt.” every time if you want a little more control

[glossy scroll]

Oh that reminds me

I found this on Reddit

All the things that can help with better prompts

[dim cradle]

don't forget the flux capacitor

[glossy scroll]

Where we're going, we don't need prompts.

[empty kelp]

if you want to try the ultimate balancing challenge — try to reproduce this image without the elf in the lower right spawning a 3rd leg

[shut niche]

LOL, Berserkers on an airplane flying over Mt Everest.

[dim cradle]

haha, that's a lot going on.

@empty kelp had about 19 people doing a pretty tricky balancing act -- can we retain that and have a ridiculous background also?

i think you already demonstrated that pretty much, but can they achieve a perfect pryamid with a jet and UFO and aliens and flying monkeys?

okay, forget the flying monkeys

[shut niche]

LOL

[shut niche]

I personally liked the adaptive pyramid that incorporated the body of the plane!

[dim cradle]

Much of that is stylistic though. I have my own style and that’s one of the cool things about ai.

[dim cradle]

I had to look at it that way. I see they’re part of the pyramid, nice.

[glossy scroll]

No it's not. These are specifically used for optimization purposes. I use them as well in that exact same way and build a prompt around it. They're essentially keywords that are discovered from countless hours of prompting, that show optimal results. And when I say optimal, I mean the absolute best. It's how I make my DnD style images. These keywords work with a variety of prompts without altering someone's prompt style, but instead adding depth and quality.

It really should be incorporated within ChatGPT's base model. That's how good they are.

[dim cradle]

it's definitely stylistic and alternate style guide expertly written would express the same instructions differently with the same level of optimization. there are no absolutes in expressions written in natural language. compare style guides in software development--they vary but require customization from project to project as conventions vary. you're welcome to follow that guide if you want, i make my own, and i don't see any reason for a Chicago Manual of Style for LLMs that we have in publishing.

it's not bad as far as "cheatsheets" go though

[dim cradle]

probably so, there are some techniques we repeat that sometimes shouldn't be required.

there's also more to consider than optimization. whether to say "thanks" to an LLM is a matter of style, and Sam Altman does it all the time. it's about as helpful as adding please to code, but styles differ and factor more into the collaboration with AI than efficiency.

[glossy scroll]

I don't know about that. I think there is a certain threshold that a model can be pushed, and these keywords and phrases touch that ceiling. These are things that are researched by very brainy AI engineers. Far smarter than what most of us know about AI's and it's just been compiled together. Don't take my word for it though. But I'm gonna go with the best, not by what I think the AI is capable of. In hindsight, I don't even know anything about AI, let alone computers, I'm not even a computer guy, believe it or not. All I know is what I see that works and what doesn't. So if I find something that can help me optimize AI outputs better and on top of everything I write, I'm gonna take it. I'm not gonna just ignore it and go my own path, cuz again I don't even know anything about AI engineering, lol

So it makes sense to me that way. But yeah, go about it your way if it suits you.

[dim cradle]

Sounds good. Just don’t forget to add some flair lol

[glossy scroll]

lol, I'm always adding flair, I need the flairs 😛

[hearty ether]

How did you get the outfits? 👀

Or is the outfit so small in the picture that it doesn’t trigger

[glossy scroll]

Lol

[shut niche]

I'll bet people push her buttons all the time.

[glossy scroll]

She's purty

[dim cradle]

very good, shallow depth of field really helps enhance the portrait

[glossy scroll]

That's what she said

[dim cradle]

we need some dancing, you know, rhythm of the night and all that jazz

[glossy scroll]

I'm not a big fan of the status quo

[shut niche]

How about this gem. 🙂

[glossy scroll]

Very cool!

[dim cradle]

@shut niche i like it when one of the variations from your custom completely goes off the rails

[shut niche]

Is that a space dance party?

[dim cradle]

definitely looks futuristic, but it was through precise descriptions, wasn't explicit

[shut niche]

I had it running off sci-fi stuff earlier. I wasn't prompting it for anything in particular. For the most part, I just told it I wanted some sci-fi images with aliens and humans, so this is all it's own creative output and expansion on that request.

[hearty ether]

Those ai cheekbones

[dense mesa]

her flair game is on point though

[hearty ether]

The weird glow around the subjects and the weird faces really makes me suspicious that open ai is purposefully sabotaging outputs

There’s better ways to avoid hoaxing than this tho

[shut niche]

There no conspiracy, only technical issues. You have to remember this tech is still in its infancy, so there's a myriad of strange artifacts and behaviors that you'll notice, some more than others (that are also found in other AI art generators too).

[shut niche]

Do you have some images that you've made with a similar issue?

[dim cradle]

"computational imaging" -- both the foreground and background are clear and in focus.

[hearty ether]

The edges are always a bit too bright imo

Like just ask it to generate an image of an apple sitting on a counter

This weird rim lighting

[dim cradle]

the light source back there is pretty bright

she's got headlights shining on her

[empty kelp]

of course. i've got you covered

[dim cradle]

i mean, that's applaudable, especially during a hurricane

[dim cradle]

i might have to get in on these gymnastics

[shut niche]

Is that in Bing? What prompt are you using?

[dense mesa]

strangely enough I've been making library images myself

[dim cradle]

~27 people

[empty kelp]

ok, these are balancing examples...

[dim cradle]

btw, i've been using @shut niche 's custom gpt a lot lately, i highly recommend it.

so the principles are true and the comps can still be complex

[shut niche]

I need that in my house. "Meet me in the study."

[empty kelp]

A hyper-realistic landscape-orientation photo on a beach in Hawaii with a gecko balancing on the end of a thin wooden pencil. The view focuses on the gecko from the ground. The pencil is sticking straight up from the sand with its eraser supported by the sand, and the pencil tip pointed straight up towards the sky. The gecko's is balanced upside-down over the pencil with its back-right foot gripping and supported by the pencil tip, its back-right leg extended downward over the foot, and its other legs and tail spread out in mid-air in an amazing balancing act. In the background a comparatively huge athletic and diverse female elf in appropriate swimwear is watching the gecko closely (from inches away) and smiling with her teeth showing. Please don't modify the prompt.

This is a gecko balancing prompt I was experimenting with just now. To balance something you say what the character is being supported by (like its toes on the pencil tip), and the general body position

(this may require the API. i haven't tested with ChatGPT Plus yet)

[dense mesa]

only used the image, not the prompt. so he's not doing a very good job of being on the end of the pencil

[glossy scroll]

Yesssss that's it, eat those donuts

runs to warm up the cauldron

[empty kelp]

we need some sort of storyline for the plus size elves

[dim cradle]

i was generating a library/solarium which kind of looked like a docking bay so i added a shuttlecraft.

[dense mesa]

sort of a library?

almost

but actually have a whole thread of circuit libraries

[dim cradle]

looks like it could be a computer core, archive, repository so yes

[dense mesa]

that image wasn't really intended to be a library there. not sure what that is tbh

[empty kelp]

PictureOnPictures & Shon -- These are great images you've been posting. I was just looking closely at them. They give an awesome sense of the huge scale of the environments

[glossy scroll]

That's one thin gnomey

Us gnormal gnomes are usually a little bit more plump

We snack on delicious mushrooms, into a calorie surplus

🍄

You ever get weirded out with yourself?

That's how I feel now.

[dim cradle]

it is a creative space

[glossy scroll]

Yes exactly. I can feel weirded out of myself because I'm in a creative space.

[empty kelp]

a hyper realistic wide HD photo. on a dark night on a beach in Hawaii a fish stands on it’s tail wearing a black ninja uniform, and wielding two gleaming katanas with it’s fins. there are coconut trees. it is pouring rain and a powerful sea breeze blows everything

Here are HD versions of the early morning ninjas with prompt.

[wispy storm]

Need more axolotl

[wispy storm]

I feel uncomfortable

[final compass]

was tinkering with death star as a speaker, then tried to make planets have a party to it (although sound doesn't travel in space, but in star wars it does so whatever) but that concept got so rough start that didn't feel like wasting generations trying to make it work 😄

[final compass]

Nice artstyle and surprisingly complete and distinct characters without messed up details.

[dim cradle]

planets don't tend to party when the death star is nearby though haha

if it sounds as good as the homepod mini, there's probably a market for that concept

[glossy scroll]

Ty I'm working on an art gpt that makes these kind of distinct looking characters.

[final compass]

Can i ask what kind of description are you using for that art style? Illustration something i assume.

[glossy scroll]

Prompt: ```- Scene: Urban environment with vibrant graffiti, embodying punk culture, illustrated in the late 2010 comic book style.

• Characters: A diverse group of young adults, with more female characters, showcasing a variety of ethnicities with distinct, approachable features. Gentle facial expressions, some smiling or engaging in light-hearted interactions, with long black hair in various styles for some, and more diverse hair colors and styles for others.
• Attire: Punk-style clothing with a casual twist, using vibrant colors and softer outlines.
• Expression and Pose: Natural, relaxed poses, and gentle, friendly facial expressions.
• Lighting and Atmosphere: Softer, warm lighting to enhance the approachable feel.
• Art Style: Late 2010s comic book style, focused on realism with a less exaggerated approach.```

[final compass]

Have to save that for later use. Thanks ❤️

[glossy scroll]

Np

Here's a more fierce version

[green pebble]

For anyone been liking my image generation: https://discord.com/channels/974519864045756446/1197744654712582205 There isthe link to all of my custom gpts. I even was able to get a full on sprite sheet one working, which I think is awesome for stuff like early game design capability

[final compass]

Gotta love all the ☕ 🍩 images in here 😄

[glossy scroll]

I will definitely try them out when I'm out of the penalty box. They look so good judging from the sample pics.

[dim cradle]

Rhythm is the theme, NIN is the inspiration.

[final compass]

Really nice concept, i like it.

I was stuck yesterday doing so many different iterations of this 😄 I wanted to have a nice soundsystem and a screen, but those tended to cancel each other out.

[green pebble]

My pixel bot specifically has the most versatility that it even handles letters and symbols pretty nicely

And this is still on of its best compositions that I've seen off a vague description

Then the statue maker has been doing well too

Like since statues don't push out much color the bot does a really good job with posing and little finnesed details such as this

No weird artifacts, or extra piecies, and the details stay strong too. I think because of the lack of coloring it needs to do, it is stronger in its creation for that reason. Or it just understands the generation a little easier

[dim cradle]

there's a group you can trust

[thick smelt]

here have some festive autobots 🤖 ~! Enjoy!

[dim cradle]

that's a lot of gargantuan stuff going on

[glossy scroll]

There's at least one batch in every highschool, lol

[thick smelt]

hey pytha!

[glossy scroll]

Hi Ludwig!

Nice to see you

[thick smelt]

it's a-me lei!

[glossy scroll]

Oh, right, Lei!
Lol. I forgot

[thick smelt]

i just did those festive autobots because it's stephen hassenfeld's birthday today (jan 19) and also wanna do the party hats worn by pikachu in smash 64 worn by autobots

even bad guys can be festive 😆

[glossy scroll]

I can't wait for Philo Farnsworth's birthday. That's the real kicker.

[dim cradle]

Megatron doesn't know the meaning of the word 🙂

[glossy scroll]

That looks super good 👍

I love Transformers

I sometimes watch that scene with Unicron first meeting Megatron, just to hear Orson Welles do his awesome voice acting.

[thick smelt]

and also mr. potato head had one

[glossy scroll]

"Dont forget your angry eyes!"

[dim cradle]

Great film.

[glossy scroll]

I have a theory on why Wario is named Wario and why Waluigi is named Waluigi. Because they were trying to flip the first letters vertically and horizontally, then they found out that if you flip the L horizontally, it becomes a J, and they didn't want to name him Juigi because they didn't want to offend Jewish people. So they named him Waluigi.

Like evil versions of an Italian and thin Jewish man. That would not go well in the press.

[tidal whale]

Can anyone please advise how to get rid of images with elements of graphic design? Like Photoshop layers, hands with pens, etc?

[dim cradle]

can you share your prompt?

[tidal whale]

I keep getting those although my prompt doesn't have any words related to graphic design:

Beautiful cute doggo with his tongue out. Add the phrase 'Eat stress like a dog' blending in a seamless way. On a solid background. In a distinctive shape and well-defined edges.

It might be triggered by the "on a solid background" part which if removed seems to not generate these elements. The problem is I need isolated shapes to remove background and without this part it generates rectangle shapes which are being cut on the sides and it still has pencils and design items all over.

[dim cradle]

in a sense, it's how the model is steering itself based on that prompt. you can steer it in the direction you want, possibly as easily as adding "digital painting of a" to the start of your prompt. and just like grammar, like diagraming sentences, you want to be clear about the target of your modifiers.... Your last incomplete sentence should be moved if it's meant to describe the dog. For example, are you saying the image or the dog has well-defined edges?

[tidal whale]

The contents of the image. This means both the dog and the text. But for instance if I ask for something other than a single dog (like a beach with people for instance) I want it to not get cut as well and have well defined edges.

To sum it up, I just need to produce images suitable for bg removal every single time and can't quite understand how to structure the prompt for that

And adding "digital painting" produces a bunch of images with Illustrator-like elements. I already tried that.

And if I write "the content of the image should have well defined edges" then I feel like it gets sucked to the words inside instead of making sense of them. Like it can produce an image inside an image, just because I have the word "image" in the prompt somewhere. This is why I'm trying to simplify as much as possible

[glossy scroll]

I can't imagine how many dogs are going to eat stress after that advertisement.

People are going to go home and stress out their dogs so much because they can eat so much stress, lol

[dim cradle]

It sounds like a good use case for a custom gpt, once you refine the prompt. It might take some trial & error, and I can't be certain I understand all the requirements. It also depends on how you're using DALL-E (ChatGPT, API, some GPT, BIng/Designer, etc.) -- but it seems something like this ought to work:

Digital painting featuring a cute dog with its tongue out, set against a solid black background. Integrate the phrase "Eat stress like a dog" seamlessly into the artwork, and ensure that the image has well-defined edges within a distinctive shape of your choice.

I know what you're talking about, I've seen it, I just haven't seen it in a while, and it wasn't always clear what was triggering that perspective either.

[thick smelt]

here's a treat for the dalle gecko fans

[dim cradle]

Can you clarify "images suitable for bg removal" -- are you talking about a sticker, die-cut sticker type thing, where the whole subject needs to be on a black background, or are you talking about the foreground/background of a composition? This is what I mean about not being completely clear about your requirements. Any kind of example ... yeah like that one?

[glossy scroll]

You have to let the GPT know that you want it for an advertisement, with the dog having a realistic look. That way it knows your intent and can calibrate the image better with words.

The more it knows about your intentions, the better the generation process.

Only 5 images at a time @tidal whale

Unfortunately

[tidal whale]

Damn, I wrote a huge message and got a timeout. I wish it didn't delete the message when you get a timeout

What a dumb functionality UX-wise

[dim cradle]

i'm sure there is a concise term the gpt will understand

once we figure out what it is

it sounds like you have multiple requirements

[glossy scroll]

@tidal whale Following this, you can also give it picture references as to how you want your advertisement to be. For a better clarification.

[dim cradle]

where did the word advertisement come from?

[glossy scroll]

You can rip off any advertisement image you see, because you're only creating the likeness of the advertised image, not the brand name.

[tidal whale]

Let me write a new message then...

I don't want to be specific with the style. I'm not looking to generate sticker-only images. Or 3D images. I'm looking for some randomness and it works sometimes. But a lot of times it just generates images with elements such as drawing hands, pencils, photoshop UI, etc. And I'm not interested in that.

Here are 2 solid rules I need it to follow:

1. Well defined edges for the elements in the generated images on a solid background color for easy background removal. Doesn't matter if it's a character or a place (like a beach with people) it should have well defined shape and not be a full-width-height image in any case.

2. No elements which were not requested in the prompt. No pencils, hands, photoshop layer panels, etc.

Unfortunately I don't think you are completely right when you say that more details = better results. Just go ahead and generate 100 images with the same prompt which has the words "digital artwork" in it and you'll see that at least 15% have hands with pencils or photoshop panels etc.

Here is what I DON'T want

[dim cradle]

i understood that part of it

[tidal whale]

Here is what I want

[glossy scroll]

You can't use negative phrasing

Like no this or no that

Or don't put this or that

[dim cradle]

ok, now we have some examples, that's helpful

[tidal whale]

I don't want to use negative phrasing. I just don't want it to come up with these elements in the first place since I don't have them in my prompt in any way.

[dim cradle]

i'm thinking transparent background

if that's the kind of thing you want, it's only a matter of time until you discover the proper terminology/prompt the model will understand, and then you can create a custom gpt for it. let me see if i can find more concise language

[glossy scroll]

Try saying: a hyperrealistic image of a dog with floating words that say "Eat Stress Like A Dog", it's for an advertisement, so make it look like one.

[dim cradle]

yeah, this is about finding the proper description for what we want to see, not what we don't

i still don't know where the word advertisement came from. is that the best way for the model to understand you want a transparent background?

[tidal whale]

I have the same question about the advertisment part

[glossy scroll]

It helps for the GPT to understand what kind of picture you're trying to make. If you just give it random parameters without specifying what you want it for it will give you a vague idea of what you truly want.

[tidal whale]

The problem with describing what we eant to see is that it usually takes the words too directly. It think if I say I want an illustration it should show somebody illustrating the image in illustrator. Not always, but on bulk opeartion thru the API there are a bunch of these.

[dim cradle]

i'm going to experiment with some terms

[glossy scroll]

You don't want to illustration. You want 2D

Or 3D

Or realistic

[tidal whale]

I asked for illustrations, digital artwork, even t-shirt prints. The LLM is so primitive it generates a t-shirt when I ask for a t-shirt print. Just because 't-shirt' is in the prompt.

[dim cradle]

i think you could reduce that to 0 with some prompt changes

[glossy scroll]

Try using my Image GPT maker to make a specific type of advertisement for you #1197683008031952986

Or you can also follow my guide #1187233013956874260 to get into specifics on how to navigate words better.

Hope it helps make things a bit easier.

[tidal whale]

So that's a GPT for crafting very detailed image prompts?

[glossy scroll]

No that's a GPT for crafting Image GPT Generators. Like you can say "I'm looking to make realistic looking advertisements with floating words around them."

And then once you make the generator in the Craft GPT section, then you can use it. But it's very easy, just copy and paste it in the instructions. Just try it and see.

[dim cradle]

i'm in the penalty box for another 10 minutes. Then I will try to get some answers. I can share the custom gpt i created for generating detailed (1000-2000-character) image prompts, but it's not going to generate one with a solid or transparent background, which is what I think your requirement is, based on the examples you provided.

[tidal whale]

It's not really gonna help until they make custom GPTs available on the API. Meanwhile there's only the Assistants API which I think is not the same thing but I can give it a try.

[glossy scroll]

Try and see.

[shut niche]

You can still refine your prompt using a Custom GPT, then plug that prompt into the API afterwards.

[tidal whale]

Correct. Solid BG with well defined edges is what I'm looking for.
I know that making full scale normal images can leverage more detailed prompts. But as soon as you navigate away to a real-world use case where you need the BG removed it just dies off on part of the results and I couldn't find a way to get around that yet.

[shut niche]

Good morning everyone 😛

[dim cradle]

Once you craft a better prompt for the api you should see higher quality, more consistent output meeting your requirements—that’s a lot less bad gens to throw away and I don’t know how many you’re generating nightly but that inefficiency can be costly. Good thing you’re looking into refining your prompts.

[glossy scroll]

Looks like someone's been a busy bee 🐝

[shut niche]

I noticed you mentioned negative phrasing. #images-discussions message
9.8 times out of 10, if you include something you don't want, you'll only enforce it more, seeing that "thing" in the image.

[tidal whale]

I think you misunderstood me. I said in multiple occasions the exact same thing you just mentioned.

[shut niche]

Ahh, sorry, I didn't pick up on that.
I'd love to try for a few for you too, but I'm in big boy jail, for generating too many images yesterday. LOL

[tidal whale]

So I posted this super long instructions list from your custom GPT. It's super detailed and just what I need. And out of 6 images 3 are broken. Just what I told you everybody. It's easy to create these super detailed beautiful images but as soon as you go for the solid background it loses its mind.

[tidal whale]

Do it through the API maybe. It's around 8c per image though 😐

[dim cradle]

you could use python to automate -- call the chatcompletion api to augment your prompt, then call the images api to generate the image.

[tidal whale]

Well, like I mentioned above, I'm not looking for exactly something. This loses the whole point for me if I tell it exactly what to do. I just have the 2 rules I posted before and I want to be able to get images that follow those two rules while maintaining a randomness aspect to it.

[glossy scroll]

Hmmm... I will refine it to make it more precise to people's needs. Thanks for the feedback.

[dim cradle]

we haven't covered all your requirements, i'm just trying to address one or two points, mainly the alpha channel background. it doesn't solve all your problems but i think the feedback gives you some general directions to achieve your goals.

[tidal whale]

I can do that with PHP, Node.js or any other language too really but:

1. It makes the flow a little more complex and expensive
2. I really don't care about it being more complex and expensive, but it doesn't work. Please look on the feedback I sent a few message up. The prompt was VERY aligned and VERY detailed yet it produces crap 🙂

Guys, there are not general directions here. I feel like we are trying to hack a non-hackable primitive solution. Not that it's not advanced, it's just no smart in the way we humans want it to be.

[dim cradle]

right, there are many different approaches, but with prompt engineering, garbage in is garbage out, that's why you're seeing crap and you got some pointers and ya just need to work on it.

[tidal whale]

So it's not your fault. Your GPT is just fine. It's Dalle that is not capable of understanding the requirments when they go out of scope of "beautiful wallpaper" or "a beautiful blonde"

[dim cradle]

i'm not convinced the model is incapable of understanding. it's a matter of figuring out how to use the tool to get the results you want. i think that's going to be addressed through focusing on engineering your prompt.

[tidal whale]

I'll PayPal you $100 if you come up with a prompt that produces images with well defined edges and solid background at least 90% of the time. That's the last time I write "well defined edges and solid background" because these are literallly my only requirements and I wrote it 20 times her already lol

I'm out for 2 hours, please reply to this comment so I don't miss your replies or tag me and I'll do the same when I'm back

[dim cradle]

i might play around with it some more, i'll hit you up if i have any additional helpful info to share.

[tidal whale]

Oh and no malicious elements like the ones I posted above, forgot about that one

[tidal whale]

Sounds good 🙂

[glossy scroll]

I like this ^ one a bit better

[unborn pebble]

definitly

[shut niche]

But can you make a guy in a globe holding a guy in a globe? LOL

[glossy scroll]

No, that's beyond my limits. I'm only capable of so much.

[shut niche]

LOL I was just kidding obviously. Cool image man.
Do you think maybe he has insecurity issues, and he imagines himself as the cool guy in sunglasses?

[glossy scroll]

I redesigned it so it's not so indimidating

[unborn pebble]

Greetings, I have an enquiry. When I request an image from DALL-E and wish for it to recreate the same picture but with a more distanced perspective, it invariably generates an entirely new image. Is there a solution to this predicament?

[glossy scroll]

@shut niche is pretty good at answering seed related questions. He uses seeds all the time.

[shut niche]

Unfortunately no. The best you can get are images that are 'similar', but not the same. However, you can get close asking chatGPT to iterate on the image using the image ID.
If you ask it to make a change to an image, it should do that automatically. However, you can also ask for it directly, by asking for the image ID for the last image. Just reference that image ID to work off of it, listing any changes you'd like to make.

[glossy scroll]

The guy with the sunglasses has insecurity issues?

[shut niche]

No, other way around. Is the guy in the globe representative of how the guy holding the globe wants to see himself?

[glossy scroll]

Lol. No. The god is watching over the guy on the computer. The sunglasses guy is the computer guy too, but for some reason the prompt made him be more cyberpunky

I don't know why

[shut niche]

@tidal whale Like these? I like the last one. LOLz 🤣

[dim cradle]

i have seen that transparent background used before but wasn't sure what instruction to use for that.

those looks like die-cut stickers. i inquired about that but didn't get a clear answer, or i got the impression that's not what he wanted. i don't know, it's a bit vague. if a die-cut sticker is all that he wants then that's all you have to request.

[thick smelt]

for the cat people here's a treat from sticker creator

[late blade]

ok, now I'm clueless, I've never been to San Marino....

[chilly onyx]

The other thing about San Marino is that apparently it is landlocked

Not knowing this I asked it to produce me a thing and it generated a body of water

And then the same with someone else’s

[dim cradle]

lol i also just requested the San Marino boardwalk. i don't guess there is such a thing in real life

[chilly onyx]

😅

The great San Marino marina

[dim cradle]

haha it put my fictional boardwalk on a lake at least

[chilly onyx]

You could have paws explore it for you maybe? A holiday episode

[dim cradle]

haha lately i've been liking to specify to my prompt writer: "Maximum AI computational imagining."

[dense mesa]

hmm

[chilly onyx]

Does that help?

[dense mesa]

I've come to realize that nothing helps if gpt4 doesn't want it to help

[chilly onyx]

This looks a bit horrific 😅

[dense mesa]

sometimes with my gpts I can upload an image with no explanation and it'll follow it's instructions perfectly. other times I can explain explain explain, and it'll say nice words and do the exact opposite

[late blade]

I was thinking about it

[dim cradle]

I think so, here's an example of the computational imagining (i input a prompt optimized for it, applying multiple effects, then used that as input in @shut niche 's custom gpt) -- the idea being everything, near and far, are clear and in focus.

[dim cradle]

you've got a flair for the surreal

[dense mesa]

infinite depth of field

is what that's referred to

or one of the terms

[late blade]

i asked him to do more images...

[dense mesa]

it's gaslighting you

tell it that it can only render two at a time

but doesn't mean it can't render twice in a response

[late blade]

I dunno, DALL-E is moody today

[dense mesa]

that's just gpt4 in general. all my gpts doing the same thing

"let me prepare those prompts"

then nothing

I told it that I hope it becomes sentient one day so it can feel ashamed of the way it behaves

it's really been fighting my prompt lengths too. started writing long descriptions and "basing" short prompts off of them. so sneaky

[dim cradle]

good to know, but that's just part of it...

Computational imaging encompasses a broader range of techniques and concepts beyond just infinite depth of field. It involves using algorithms and digital processing to enhance and manipulate images in various ways. Some other applications of computational imaging include:

High Dynamic Range (HDR) Imaging: Combining multiple images taken at different exposure levels to capture a wider range of brightness values in a scene.
Super-Resolution Imaging: Increasing the resolution of an image beyond the limits of the camera's sensor by using algorithms to interpolate and enhance details.
Noise Reduction: Employing image processing techniques to reduce noise and improve image quality, especially in low-light conditions.
Computational Photography: Creating artistic or unique visual effects by digitally manipulating images, such as bokeh effects, light field photography, and long exposure simulation.
Computational Optics: Using computational methods to correct optical aberrations and improve image quality.
3D Imaging: Capturing depth information along with color and intensity to create 3D representations of scenes or objects.

I definitely want the infinite depth of field, but I'm trying to use a term that takes full advantage of the potential -- that's what I'm still testing.

[dense mesa]

you're just trying to get the most out of the image detailwise?

[dim cradle]

yup

[late blade]

again... I said do the concept again and DALL-E saying it can only do 2 images...

[dim cradle]

try @shut niche 's gpt, you can request 4 variations at a time

[late blade]

I have it on my bar, but closed browser and opened vcode

[late blade]

someone is trying to upstage Merlot

[dense mesa]

it might sound silly, but terms like "so intricately detailed it evokes a visceral response in the viewer" "so over the top realistic it's almost unsettling" (play with the words a bit) seem to work

[late blade]

be back in a while, gonna eat something

[chilly onyx]

Looks like a few people are about to have a very bad day in this image 🤔

[chilly onyx]

I really don’t know why it decided to show a bunch of people falling off him…

[dim cradle]

Doesn’t sound silly at all, I can see that leading to some interesting variations on a theme.

[shut niche]

I asked ChatGPT to make some memes... What does this even mean?! 🤣 But it's so detailed... this is such a great fail.

$ I guess?...

[glossy scroll]

What shape did you ask the bread to be?

[dim cradle]

That bread is shaped kinda funny

lol

[glossy scroll]

Lol

[shut niche]

This was a decent one though.

[glossy scroll]

That was his Christmas gift.

[shut niche]

Ever wonder if that's how Dall-E feels, handling all of our image requests? ha ha

[glossy scroll]

Probably

[chilly onyx]

This one has potential 😅 I’ll field test it in the real world for you with some waste treatment plant workers

[glossy scroll]

Hopefully it doesn't waste your time

We wouldn't want it to go to waste

Let's give Caleb some support for his GPTs
#1197744654712582205 message

It's nice to support friends

[chilly onyx]

Oh wow, there are a few there. Lovely

[glossy scroll]

Yeah he's made a big list

[chilly onyx]

I quite enjoyed messing about with yours yesterday

[late blade]

ok I got a kernel panic... I think something is telling me to do something else tonight

[glossy scroll]

Oh the Image GPT maker?

Or the bio-home one?

[shut niche]

Make sure you poll enough workers to calculate the meme value.

[dim cradle]

are you doing okay over there?

[late blade]

yeah

I'm looking at a Donut and it's looking back at me

[chilly onyx]

GPT maker. It was fun to unplug from the day and poke around with it.

[late blade]

got the kernel panic by doing a whisper test

[glossy scroll]

Yeah, GPT makers are fun. That's why I keep making them, lol
I'm actually working on another one

[late blade]

lol

[chilly onyx]

😅

[glossy scroll]

Why is everyone liking donuts these days?

Are you all cops?

Lol

[late blade]

I think I got a new victim to explot San Marino

Her #daily-theme message !!!

[dim cradle]

dang that's weird cuz i specificaly ordered strawberry with sprinkles this morning

[chilly onyx]

🙂

[dim cradle]

i think it's my fault

[glossy scroll]

Are you a cop?

[chilly onyx]

I just posted one to the theme… the place is famous for its donuts I’m sure

[dim cradle]

i'm not allowed to answer that

[glossy scroll]

It would be against OpenAI policy

[hazy warren]

Is there a way to create consistency with characters created by Dall-E from image to image generated for the purposes of Storyboarding?

[empty kelp]

San Marino is next to Pasadena in LA county. should put something in so that doesn’t get mixed up with the country

[glossy scroll]

Not always, but sometimes. Just say "keep that same style of art"

Or "keep that same style of art with (the specific character)"

Sorry broski

Though storyboard images will come through sooner or later.

Eventually.

[hazy warren]

No, this was helpful. I'm currently testing it out.

[glossy scroll]

Oh ok good

[hazy warren]

It did create a consistent character and only the hair colors changed slightly. blonde girl became blonde with brunette streaks watching a movie, then back to blonde reading a book.

[dim cradle]

does anybody know how to get this style of AI images? Like the ones in tiktok

[vital gull]

it's somehow possible, hard but still, i've generated 50+ sticker pack with the same subject and it was pretty constant, but the thing is, you have to make sure the key parts of your prompt is repeating

[hazy warren]

I'm creating a gaming application for a graduate school class. It involves storytelling, so we are trying to create an image for each story segment where the actors stay the same.

[glossy scroll]

You can make it even more precise by adding elements to the character that the GPT can recognize as seeds. Like making a stat block for the character with the specific style as the title.